22.3.1 Temporal Expression Recognition
Temporal expressions are those that refer to absolute points in time, relative times, durations and sets of these. Absolute temporal expressions are those that can be mapped directly to calendar dates, times of day, or both. Relative temporal expressions map to particular times via some other reference point (as in a week from last Tuesday.) Finally, durations denote spans of time at varying levels of granularity (seconds, minutes, days, weeks, centuries etc.) Fig. 22.17 provides some sample temporal expressions in each of these categories.
Syntactically, temporal expressions are syntactic constructions that have temporal lexical triggers as their heads. In the annotation scheme in widest use, lex
| Absolute | Relative | Durations |
| April 24, 1916\nThe summer of '77\n10:15 AM\nThe 3rd quarter of 2006 | yesterday\nnext semester\ntwo weeks from yesterday\nlast quarter | four hours\nthree weeks\nsix days\nthe last three quarters |
| Category | Examples |
| Noun | morning, noon, night, winter, dusk, dawn |
| Proper Noun | January, Monday, Ides, Easter, Rosh Hashana, Ramadan, Tet |
| Adjective | recent, past, annual, former |
| Adverb | hourly, daily, monthly, yearly |
| Figure 22.18 | Examples of temporal lexical triggers. |
ical triggers can be nouns, proper nouns, adjectives, and adverbs; full temporal expression consist of their phrasal projections: noun phrases, adjective phrases and adverbial phrases. Figure 22.18 provides examples of lexical triggers from these categories.
The annotation scheme in a widest use is derived from the TIDES standard (Ferro et al., 2005). The approach presented here is based on the TimeML effort (Pustojovsky et al., 2005). TimeML provides an XML tag, TIMEX3, along with various attributes to that tag, for annotating temporal expressions. The following example illustrates the basic use of this scheme (ignoring the additional attributes, which we'll discuss as needed later in Sec. 22.3.2).
$$ \mathrm{~A~f a r e~i n c r e a s e~i n i t i a t e d~}\mathrm{
Corp's United Airlines was matched by competitors over
The temporal expression recognition task consists of finding the start and end of all of the text spans that correspond to such temporal expressions. Although there are myriad ways to compose time expressions in English, the set of temporal trigger terms is, for all practical purposes, static and the set of constructions used to generate temporal phrases is quite conventionalized. These facts suggest that any of the major approaches to finding and classifying text spans that we've already studied should be successful. The following three approaches have all been successfully employed in recent evaluations.
- Rule-based systems based on partial parsing or chunking
• Statistical sequence classifiers based on standard token-by-token IOB encod-
ing
• Constituent-based classification as used in semantic role labeling
Rule-based approaches to temporal expression recognition use cascades of automata to recognize patterns at increasing levels of complexity. Since temporal expressions are limited to a fixed set of standard syntactic categories, most of these systems make use of pattern-based methods for recognizing syntactic chunks. That is, tokens are first part-of-speech tagged and then larger and larger chunks are recognized using the results from previous stages. The only difference from the usual partial parsing approaches is the fact that temporal expressions must contain temporal lexical triggers. Patterns must, therefore, contain either specific trigger words (e.g. February), or patterns representing classes (e.g. MONTH). Fig. 22.19 illustrates this approach with a small representative fragment from a rule-based system written in Perl.
Sequence labeling approaches follow exactly the same scheme introduced in Ch. 13 for syntactic chunking. The three tags I, O and B are used to mark tokens that are either inside, outside or begin a temporal expression, as delimited by TIMEX3 tags. Example 22.3.1 would be labeled as follows in this scheme.
A fare increase initiated last week by UAL Corp's...
As expected, features are extracted from the context surrounding a token to be tagged and a statistical sequence labeler is trained using those features. As with syntactic chunking and named entity recognition, any of the usual statistical sequence methods can be applied. Fig. 22.20 lists the standard features used in the machine learning-based approach to temporal tagging.
Constituent-based methods combine aspects of both chunking and token-by-token labeling. In this approach, a complete constituent parse is produced by automatic means. The nodes in the resulting tree are then classified, one by one, as to whether they contain a temporal expression or not. This task is accomplished by training a binary classifier with annotated training data, using many of the same features employed in IOB-style training. This approach separates the classification problem from the segmentation problem by assigning the segmentation problem to the syntactic parser. The motivation for this choice was mentioned earlier; in currently available training materials, temporal expressions are limited to syntactic constituents from one of a fixed set of syntactic categories. Therefore, it makes sense to allow a syntactic parser to solve the segmentation part of the problem.
In standard evaluations, temporal expression recognizers are evaluated using the usual recall, precision and F-measures. In recent evaluations, both rule-based and statistical systems achieve about the same level of performance, with the best systems reaching an F-measure of around 0.87 on a strict exact match criteria. On a
yesterday/today/tomorrow
$string = ~ s/(($OT+(earlyearlier|later?)($CT+\s+)?((($OT+the$CT+\s+)?$OT+day$CT+\s+
$OT+(before|after)$CT+\s+)?$OT+$TERelDayExpr$CT+(\s+$OT+(morning|afternoon|evening|night)
$CT+)?)/
$string = ~ s/(($OT+\w+$CT+\s+)
this/that (morning/afternoon/evening/night)
$string = ~ s/(($OT+(earlyearlier|later?)($CT+\s+)?$OT+(this|that|every|the$CT+\s+
$OT+(next|previous|following))($CT+\s*$OT+(morning|afternoon|evening|night)
$CT+(\s+$OT+thereafter$CT+)?)/
Figure 22.19 Fragment of Perl code from MITRE's TempEx temporal tagging system.
| Feature | Explanation |
| Token Tokens in window Shape POS Chunk tags Lexical triggers | The target token to be labeledBag of tokens in the window around a targetCharacter shape featuresParts of speech of target and window wordsBase-phrase chunk tag for target and words in a windowPresence in a list of temporal terms |
| Figure 22.20 Typical features used to train IOB style temporal expression taggers. | |
looser criterion based on overlap with gold standard temporal expressions, the best systems reach an F-measure of .94. $ ^{4} $
The major difficulties for all of these approaches are achieving reasonable coverage, correctly identifying the extent of temporal expressions and dealing with expressions that trigger false positives. The problem of false positives arises from the use of temporal trigger words as parts of proper names. For example, all of the following examples are likely to cause false positives for either rule-based or statistical taggers.
(22.19) 1984 tells the story of Winston Smith and his degradation by the totalitarian state in which he lives.
(22.20) Edge is set to join Bono onstage to perform U2's classic Sunday Bloody Sunday.
(22.21) Black September tried to detonate three car bombs in New York City in March 1973.
July 2, 2007 A fare increase initiated